Internship work at Lennar

This is the documentation of my work at Lennar Corporation. Leveraging LLM to interpret data to manage construction process.

• Technology Stack: Python, FastAPI, OpenAI APIs, FAISS, LangChain
• Timeline: June 2025 - August 2025
• Role: Applied AI Engineer Intern
• Job description:
  • Architected scalable FastAPI microservices with object-oriented design supporting multimodal AI workflows, processing 1000+ images with hybrid search (semantic + keyword + reranking) in 240 seconds through async batch processing and efficient API orchestration.
  • Built high-performance vector database infrastructure using FAISS with O(1) metadata lookup and hybrid retrieval strategies, enabling real-time semantic search across construction domain knowledge with sub-second query response times.
  • Engineered end-to-end NLP pipeline processing construction conversations into structured data reports using few-shots classification, K-means/HDBSCAN clustering, and RAG framework with LangChain integration, achieving 20k+ text embeddings in under 300 seconds.
  • Optimized ML system performance through asyncio implementation and concurrent OpenAI's API calls, delivering measurable precision/recall improvements on production-scale noisy text data while maintaining system reliability.
• Project problem statement:
  • Automated quality control extraction from construction communications: Identify QC issue, extract issue, enrich with details, output issue form
  • Image parse: keywords labeling for each site image
  • Document Image Parse: Extract structured data for each document image
  • ImageSearch that can be searched by Natural Language

Project 1: Image search

1. Overview & Data flow:
   • Input: Image parse & User query Input
   • Output: Top k related image
2. Key features:
   • Image Parser: Extracts visual features and metadata from images
   • Document Image Parser: Analyzes construction documents, QC reports, and forms
   • Image Search Engine: Enables natural language queries to find relevant images
3. Demo with synthetic data:

Project 2: Issue detect and extract

1. System Architecture:
   • Input: Single target message, supported by previous messages & external data from existing database
   • Output: Issue extracted & Evaluation report
   • Key Features: Data preprocessing, Classification & Clustering, Extraction, Enrichment
2. Sample output with synthetic data: